TabNet-Driven Interpretable Prediction of Multi-Oxide Composition in Cement Using NIR Spectroscopy

Gezhichen Li^*

• Shandong Normal University, Jinan, China

Accurate monitoring of oxide compositions is essential for ensuring the quality and performance of cement in industrial production. Conventional analytical techniques for quantifying oxides are often time-consuming, costly, and lacking real-time capability. Near-infrared (NIR) spectroscopy has emerged as a promising alternative for rapid and non-destructive cement analysis. However, traditional chemometric approaches (e.g., partial least squares regression, support vector regression) struggle to capture the highly nonlinear, high-dimensional spectral characteristics and exhibit limited interpretability. To address these challenges, this paper propose an interpretable TabNet-based multi-output regression method for predicting multiple oxide concentrations from NIR spectra. The proposed method integrates sparse feature selection with adaptive information aggregation, dynamically prioritizing the most informative spectral regions during processing. This facilitates automatic wavelength selection and accurate oxide content prediction. In addition, TabNet enhances interpretability by generating sequential attention masks that highlight chemically meaningful wavebands associated with each oxide component. Extensive experiments on two cement datasets demonstrate that TabNet consistently outperformed baseline models. This framework provides a scalable and insightful solution for spectral-based analysis in cement quality monitoring and other materials science applications. The code can be found at https://github.com/Andrew-Leopard/CementOxidePredictor.